Sri Lankan Seagrass Family Identification using Underwater Images through Image Enhancement and Deep Learning Techniques

Abstract

Seagrass ecosystems are integral to marine biodiversity, carbon sequestration, and coastal re silience, yet their monitoring remains constrained by the complexities of underwater imaging and manual species identification. Automated identification of seagrass families is important for enabling large scale, consistent, and timely monitoring, reducing reliance on labor intensive manual classification, and improving conservation and management efforts for these vital marine ecosystems. This study proposes a novel hy brid deep learning framework for the automated classification of three predominant seagrass families in Sri Lanka’s coastal waters: Hydrocharitaceae, Cymodoceaceae, and Ruppiaceae. A total of 700 underwater im ages were captured using iPhone 15 Pro Max and GoPro cameras from multiple coastal locations, including Mannar, Trincomalee, and Jaffna. Following a two stage enhancement pipeline combining Water-Net and DehazeFormer, 217 high quality images were retained, which were then augmented to 1,519 images across the three families to improve dataset diversity. Among five evaluated convolutional neural networks, VGG16, VGG19, and MobileNetV2 were selected as base learners in an ensemble model, achieving 99% classification accuracy on the test dataset. These findings demonstrate the potential of deep learning based, automated seagrass monitoring to enable scalable, consistent, and timely conservation efforts.